Research into the Architecture of CAD Based Robot Vision Systems
Abstract
The research conducted under this grant has been to investigate techniques for recognizing object that are partially hidden or occluded. It was assumed that geometrical data about the objects was available either as a by- product of the CAD process used to design the object or as the output of a training phase. There were three aspects to this work. First, recognition techniques were developed that used only a single view gray-level image. Second, complementary techniques were developed that used range images. Third, computation times for both of these techniques were studied with the intent of reducing them through the use of parallel processing. The work on gray-level images resulted in a number of new techniques to speedup recognition. These included the concept of saliency and critical point neighborhood, and a fast matching algorithm based on k-d trees. Techniques for recognizing occluded objects that can be at varying distances from the viewer were also developed. The work with range data was aimed at the recognition and pose estimation of three-dimensional industrial parts. The major subproblems addressed were: (1) segmentation and surface parameter extraction, (2) grouping regions into a set (a Convex Region Set), and (3) Convex Region Set matching. The work to develop fast parallel versions of our recognition algorithms made extensive use of a 64 processor NCUBE hypercube multiprocessor. Speedups of 30:1 over VAX 11/780 implementations were obtained for some of the low-level image processing algorithms used in the work.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 09, 1988
- Accession Number
- ADA193330
Entities
People
- R. A. Volz
- T. N. Mudge
Organizations
- University of Michigan